AI hallucination risk in audit working papers

AI hallucination risk in audit working papers shows up as a specific, narrow failure: a model asked to summarise, extract or draft something produces a fact — a figure, a date, a clause reference, a vendor name — that is fluent and plausible but not actually present in the source document, and because it reads like every other sentence in the file, nobody flags it unless the number behind it is independently checked against the source.

Where hallucination actually enters a working paper

It rarely enters through an obviously AI-written paragraph. It enters through small, specific substitutions inside otherwise correct text: a model summarising a lease agreement states a renewal date one month off; a model drafting a risk narrative cites a materiality threshold that doesn't match the actual planning memo; a model extracting invoice fields picks up a number from the wrong line item because two numbers on the page were close together. None of these look wrong on a read-through. They look wrong only against the source.

• Field extraction from scanned or photographed documents, where a model guesses at a smudged or cropped figure rather than flagging it as unreadable.
• Narrative summarisation of long client-provided documents, where the model compresses detail and occasionally invents a plausible-sounding specific to fill a gap in the summary.
• Cross-referencing claims, where a model states that a figure 'ties to' a schedule without having actually performed the comparison.
• Citing a standard, section or clause number from memory rather than from the actual document being reviewed.

Why normal review doesn't catch it

Standard working paper review is built to catch reasoning errors and omissions — a reviewer reading for whether the conclusion follows from the evidence described. A hallucinated fact doesn't fail that test, because the reasoning built on top of it is often perfectly sound; the flaw is one layer down, in whether the fact was ever actually in the source. A reviewer who doesn't re-open the source document has no way to catch it from the narrative alone.

The specific checks that catch it

1. 01Trace every AI-drafted figure back to its source before sign-offIf a model extracted or summarised a number, the reviewer opens the underlying document and checks that specific figure, not the surrounding narrative.
2. 02Require the tool to show its source span, not just its outputA tool that highlights exactly which text or cell it drew a claim from makes verification fast. A tool that only shows a clean summary makes verification expensive enough that it quietly stops happening.
3. 03Treat low-confidence extractions as flagged, not silently filled inA document that was partly illegible should produce a visible gap, not a guessed value presented with the same confidence as a clean extraction.
4. 04Keep the model out of anything that functions as the final computationReconciliations, totals and variance calculations should be produced by deterministic code, with the model confined to drafting language around a number it did not generate.